OpenAI has released a new AI large language model, dubbed GPT-4o, which is set to be the foundation for its chatbot ChatGPT.

The new model boasts enhanced reaction times and processing of images, texts, and audio, as well as increase proficiencies in languages other than English.

Its faster responses are designed to sound more human-like, even verging on flirtatious at times.

GPT-4o also meets GPT-4 Trubo on text, reasoning and coding intelligence performance, with higher performance on multilingual vision and capabilities.

The model is more conversational, as it can be interrupted and the delay between responses is on par with humans.

In demonstrations performed and provided by OpenAI, two chatbots conversed with each other in a somewhat natural manner. Still, the models are far from human – one model continued to describe the light fixtures rather than mention a new person had emerged into the room during a live take.

What sets GPT-4o apart from previous models is its training on voice mode, or the mode where a human has a conversation with the application via their voice.

Prior to GPT-4o, Voice Mode was a pipeline of three different models, one to transcribe audio to text, one to take in text and output text, and one that converts text back to audio. This looses a lot of information, such as tone, recognition of multiple speakers, and background noise, as well as different expressions like laughter or singing.

Recommended

• OpenAI Provides Glimpse into ChatGPT’s Instruction Manual
• OpenAI, Microsoft Launch $2M Fund To Battle Election Deepfakes
• Google Implements Strict Ban on Ads for Deepfake Pornography

With GPT-4o, the single new model was trained end-to-end across text, vision, and audio, meaning that all inputs and outputs are processed by the same neural network, meaning minimal loss across transcription.

The new model, however, is not without its glitches. The model did misidentify a person as a wooden surface, and at some point began to solve and equation it had not been presented with.

While the model does have work to do, OpenAI is rolling it out across ChatGPT’s free tier, with another version with 5x the response speed available to those with a paying subscription.